More
Submit build

Builds / Nvidia GB10 (unified)

DGX Spark: Inkling 2 Mini NVFP4/MXFP8 on vLLM TP1, 47.7 tok/s eval (Tech2wild)

Last updated:

Source: github.com โ†—More source: github.com โ†—More source: recipes.vllm.ai โ†—๐Ÿ“ How measured: reported by the original source; RigBase re-checked the numbers at that source
GPU(s)
Nvidia GB10 (unified)
VRAM / unified memory (GB)
128
CPU
Nvidia GB10 Grace (Arm)
Amazon ยท eBay (used) ยท Newegg
Motherboard
not stated
RAM
128 GB LPDDR5X unified per Spark (128 GB total)
PSU
not stated
Other
1 NVIDIA DGX Spark
Reported total cost
not stated
Power
not stated

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Inkling 2 MiniNVFP4/MXFP8 + DSparkvLLM TP1, 1 stream47.7not statedโšก spec-decode (DSpark (num_speculative_tokens 7))context: 262K standard tier: BF16 KV, gpu-memory-utilization 0.80 (68.89 GiB KV, 1,346,671-token pool)๐Ÿ“ 69-scenario tool eval (2Wild-Model-Eval), one request at a time, temperature 0, reasoning_effort none; decode tok/s as published in Tech2wild's README. Short-chat 72.8 tok/s is a peak and not used. Standard tier one-user chats: about 40 to 45 tok/s. Eval quality 88.4/100.
Inkling 2 MiniNVFP4/MXFP8 + DSpark, FP8 KVvLLM TP1, 1 stream47.3not statedโšก spec-decode (DSpark (num_speculative_tokens 7))context: 1M max FP8 tier: 74 GiB FP8 (E4M3) KV, 3,308,911-token pool๐Ÿ“ 69-scenario tool eval (2Wild-Model-Eval), one request at a time, temperature 0, reasoning_effort none; max FP8 tier decode tok/s as published in Tech2wild's README. Short-chat 72.8 tok/s is a peak and not used. Eval quality 87.0/100.
Inkling 2 MiniNVFP4/MXFP8 + DSpark, FP8 KVvLLM TP1, 1 stream47.21,693 @ 199,989-token prompt
est. TTFT @8k โ‰ˆ 4.8 s (calculated from measured prefill)
measured TTFT: 118 s @ 199,989-token prompt
โšก spec-decode (DSpark (num_speculative_tokens 7))context: 199,989-token prompt on the 1M max FP8 tier (74 GiB FP8 KV)๐Ÿ“ Single-request needle test at 50% depth as published in Tech2wild's README; the README says needle-row decode comes from short answers (77 tokens) and is rough. Prefill = prompt tokens / TTFT. Needle PASS.
Inkling 2 MiniNVFP4/MXFP8 + DSpark, BF16 KVvLLM TP1, 1 stream28.31,680 @ 199,989-token prompt
est. TTFT @8k โ‰ˆ 4.9 s (calculated from measured prefill)
measured TTFT: 119 s @ 199,989-token prompt
โšก spec-decode (DSpark (num_speculative_tokens 7))context: 199,989-token prompt on the 1M max BF16 tier (80 GiB BF16 KV, 2,075,828-token pool)๐Ÿ“ Single-request needle test at 50% depth as published in Tech2wild's README; the README says needle-row decode comes from short answers (149 tokens) and is rough. Prefill = prompt tokens / TTFT. Needle PASS.
Inkling 2 MiniNVFP4/MXFP8 + DSpark, FP8 KVvLLM TP1, 1 stream18.4598 @ 949,953-token prompt
est. TTFT @8k โ‰ˆ 13.7 s (calculated from measured prefill)
measured TTFT: 1,588 s @ 949,953-token prompt
โšก spec-decode (DSpark (num_speculative_tokens 7))context: 949,953-token prompt on the 1M max FP8 tier (74 GiB FP8 KV)๐Ÿ“ Single-request needle test at 50% depth as published in Tech2wild's README; the README says needle-row decode comes from short answers (91 tokens) and is rough. Prefill = prompt tokens / TTFT. Needle PASS (answered correctly at ~950K).

Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 รท prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.

Numbers by Tech2wild (tonyd2wild); Thinking Machines Lab official NVFP4/MXFP8 checkpoint; vLLM inkling-2-mini image and Inkling 2 Mini recipe (recipes.vllm.ai/thinkingmachines/Inkling-2-Mini). RigBase read the GitHub README on 2026-11-21 and checked each figure against the text. RigBase did not re-run it. Speculative decoding (MTP/DFlash/DSpark) inflates predictable text; these rows prefer prose/code over counting peaks. Recipe license: MIT.

Source: https://github.com/tonyd2wild/Inkling-2-Mini-NVFP4-DGX-Spark
https://github.com/tonyd2wild/Inkling-2-Mini-NVFP4-DGX-Spark/blob/main/README.md
https://recipes.vllm.ai/thinkingmachines/Inkling-2-Mini

Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.

Embed this result

Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds โ€” numbers stay on the build page.

Markdown

[![Measured on RigBase](https://getrigbase.com/badge/tony-dgx-spark-inkling-2-mini-nvfp4-dgx-spark.svg)](https://getrigbase.com/builds/tony-dgx-spark-inkling-2-mini-nvfp4-dgx-spark/)

HTML

<a href="https://getrigbase.com/builds/tony-dgx-spark-inkling-2-mini-nvfp4-dgx-spark/" rel="noopener"><img src="https://getrigbase.com/badge/tony-dgx-spark-inkling-2-mini-nvfp4-dgx-spark.svg" alt="Measured on RigBase" height="20"/></a>

Preview: Measured on RigBase ยท How to use

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Quick answers: Can DGX Spark (GB10) run Inkling 2 Mini?

Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).